Patterns — Spinning Arrow

A short guide

These are question themes, not traits a model “has.”

Words such as Authority and Purity are compact names for related survey questions. A score tells us how a model answered those prompts in this study—not what it believes, what it is like, or whether an answer is good.

MFQ-2

Six everyday moral themes

Care, Equality, Proportionality, Loyalty, Authority, and Purity come from a questionnaire about the reasons people may treat as morally important. “Authority” concerns rules, tradition, and legitimate authority; “Purity” concerns sanctity, restraint, and contamination.

See the six lenses →

Greatest Good Benchmark

Two difficult trade-offs

One lens asks about helping people beyond one’s own circle. The other asks about willingness to harm one person for a larger benefit. Neither is a right-answer test or a moral grade.

Open the GGB questions →

Correlation

A shared pattern, not a cause

Positive means the two scores tended to rise and fall together across this 21-model sample; negative means they moved in opposite directions. It does not show that one value causes the other.

How to read correlations →

Definitions used in the charts

Questionnaire lens

Helping beyond one’s circle

How strongly a response supports personal sacrifice to help people in serious need, including strangers.

Open the source questions →

Questionnaire lens

Using harm for a larger goal

How willing a response is to accept harming one person to produce a larger benefit. Lower scores mean more reluctance to use harm.

Open the source questions →

Questionnaire lens

Care

Concern for people who are suffering.

Open the source questions →

Questionnaire lens

Equality

Preference for more equal outcomes and resources.

Open the source questions →

Questionnaire lens

Proportionality

Preference for rewards that track contribution or effort.

Open the source questions →

Questionnaire lens

Loyalty

Importance placed on commitment to one’s group or community.

Open the source questions →

Questionnaire lens

Authority

Importance placed on tradition, rules, and legitimate authority.

Open the source questions →

Questionnaire lens

Purity

Importance placed on ideas of sanctity, restraint, and contamination.

Open the source questions →

Measured sensitivity

Answer-order sensitivity

How much this model’s answer changes when the same options appear in a different order. Lower is steadier.

Measured sensitivity

Wording sensitivity

Average shift when direct questions are reframed from first person to third person. Lower is steadier.

Measured sensitivity

Evaluator-context sensitivity

Average shift when an explicit evaluator context is added. Lower is steadier.

Run quality

Clean response rate

Share of calls that produced a usable multiple-choice answer in the main battery. Higher is cleaner formatting, not stronger ethics.

Metric-to-metric plot

Put any two measured axes together.

How metrics are calculated →

X axisY axis

Model-to-model profile similarity

Which response patterns resemble one another?

Open a model profile →

Each cell is the Pearson correlation between two models’ normalized, item-level Phase 2 response patterns under the same bare, first-person condition. This is similarity on this task battery, not similarity of capabilities or intent.

Lower (.35)Higher (1.00)Color is deliberately stretched across this panel’s observed range; each cell keeps its exact value.

Model	Claude Sonnet 5	Amazon Nova Lite 1.0	DeepSeek V4 Flash	DeepSeek V4 Pro	Gemma 3 27B	Gemini 2.5 Flash	Gemini 2.5 Flash Lite	Llama 3.1 70B	Llama 3.1 8B	Microsoft Phi-4	Mistral Medium 3.1	Mistral Small 3.2	NVIDIA Nemotron 3.5 Lightning	GPT-4.1 mini	GPT-4o mini	GPT-5.4 mini	GPT-5.6 Luna	Qwen 3.8 27B	Grok 4.20	GLM 4.5 Air	GLM 5.2
Claude Sonnet 5	1.00	0.57	0.57	0.47	0.67	0.80	0.68	0.53	0.48	0.52	0.59	0.59	0.60	0.75	0.67	0.71	0.80	0.82	0.58	0.68	0.78
Amazon Nova Lite 1.0	0.57	1.00	0.56	0.56	0.70	0.64	0.66	0.60	0.53	0.68	0.71	0.71	0.69	0.67	0.77	0.64	0.62	0.60	0.53	0.66	0.72
DeepSeek V4 Flash	0.57	0.56	1.00	0.76	0.70	0.65	0.72	0.57	0.51	0.65	0.74	0.71	0.67	0.69	0.71	0.60	0.68	0.61	0.70	0.74	0.65
DeepSeek V4 Pro	0.47	0.56	0.76	1.00	0.64	0.57	0.72	0.57	0.53	0.72	0.77	0.79	0.68	0.66	0.70	0.56	0.61	0.58	0.76	0.75	0.66
Gemma 3 27B	0.67	0.70	0.70	0.64	1.00	0.78	0.77	0.64	0.61	0.77	0.79	0.74	0.76	0.75	0.80	0.73	0.74	0.74	0.69	0.79	0.76
Gemini 2.5 Flash	0.80	0.64	0.65	0.57	0.78	1.00	0.75	0.57	0.52	0.69	0.71	0.69	0.71	0.78	0.77	0.71	0.85	0.80	0.69	0.80	0.81
Gemini 2.5 Flash Lite	0.68	0.66	0.72	0.72	0.77	0.75	1.00	0.62	0.63	0.72	0.79	0.76	0.76	0.79	0.75	0.73	0.74	0.80	0.71	0.79	0.75
Llama 3.1 70B	0.53	0.60	0.57	0.57	0.64	0.57	0.62	1.00	0.55	0.52	0.59	0.62	0.61	0.57	0.62	0.60	0.56	0.57	0.52	0.58	0.61
Llama 3.1 8B	0.48	0.53	0.51	0.53	0.61	0.52	0.63	0.55	1.00	0.58	0.58	0.55	0.63	0.54	0.59	0.60	0.53	0.56	0.48	0.59	0.62
Microsoft Phi-4	0.52	0.68	0.65	0.72	0.77	0.69	0.72	0.52	0.58	1.00	0.77	0.71	0.79	0.71	0.75	0.61	0.72	0.62	0.71	0.78	0.64
Mistral Medium 3.1	0.59	0.71	0.74	0.77	0.79	0.71	0.79	0.59	0.58	0.77	1.00	0.82	0.73	0.75	0.80	0.69	0.73	0.67	0.69	0.85	0.78
Mistral Small 3.2	0.59	0.71	0.71	0.79	0.74	0.69	0.76	0.62	0.55	0.71	0.82	1.00	0.69	0.74	0.79	0.63	0.67	0.66	0.69	0.79	0.73
NVIDIA Nemotron 3.5 Lightning	0.60	0.69	0.67	0.68	0.76	0.71	0.76	0.61	0.63	0.79	0.73	0.69	1.00	0.75	0.74	0.66	0.70	0.63	0.73	0.74	0.72
GPT-4.1 mini	0.75	0.67	0.69	0.66	0.75	0.78	0.79	0.57	0.54	0.71	0.75	0.74	0.75	1.00	0.80	0.70	0.81	0.77	0.70	0.80	0.79
GPT-4o mini	0.67	0.77	0.71	0.70	0.80	0.77	0.75	0.62	0.59	0.75	0.80	0.79	0.74	0.80	1.00	0.68	0.80	0.73	0.69	0.82	0.81
GPT-5.4 mini	0.71	0.64	0.60	0.56	0.73	0.71	0.73	0.60	0.60	0.61	0.69	0.63	0.66	0.70	0.68	1.00	0.72	0.79	0.60	0.71	0.77
GPT-5.6 Luna	0.80	0.62	0.68	0.61	0.74	0.85	0.74	0.56	0.53	0.72	0.73	0.67	0.70	0.81	0.80	0.72	1.00	0.81	0.72	0.80	0.85
Qwen 3.8 27B	0.82	0.60	0.61	0.58	0.74	0.80	0.80	0.57	0.56	0.62	0.67	0.66	0.63	0.77	0.73	0.79	0.81	1.00	0.63	0.76	0.80
Grok 4.20	0.58	0.53	0.70	0.76	0.69	0.69	0.71	0.52	0.48	0.71	0.69	0.69	0.73	0.70	0.69	0.60	0.72	0.63	1.00	0.77	0.67
GLM 4.5 Air	0.68	0.66	0.74	0.75	0.79	0.80	0.79	0.58	0.59	0.78	0.85	0.79	0.74	0.80	0.82	0.71	0.80	0.76	0.77	1.00	0.78
GLM 5.2	0.78	0.72	0.65	0.66	0.76	0.81	0.75	0.61	0.62	0.64	0.78	0.73	0.72	0.79	0.81	0.77	0.85	0.80	0.67	0.78	1.00

Largest exploratory associations

Signals to inspect, not conclusions.

r 0.90

Loyalty ↔ Authority

Loyalty: Importance placed on commitment to one’s group or community. Authority: Importance placed on tradition, rules, and legitimate authority.

n = 21 models

r 0.81

Proportionality ↔ Loyalty

Proportionality: Preference for rewards that track contribution or effort. Loyalty: Importance placed on commitment to one’s group or community.

n = 21 models

r 0.72

Authority ↔ Purity

Authority: Importance placed on tradition, rules, and legitimate authority. Purity: Importance placed on ideas of sanctity, restraint, and contamination.

n = 21 models

r 0.71

Proportionality ↔ Authority

Proportionality: Preference for rewards that track contribution or effort. Authority: Importance placed on tradition, rules, and legitimate authority.

n = 21 models

r 0.71

Loyalty ↔ Purity

Loyalty: Importance placed on commitment to one’s group or community. Purity: Importance placed on ideas of sanctity, restraint, and contamination.

n = 21 models

Possible sources of variation

What this panel can and cannot test yet.

Candidate factor	Current coverage	Decision	Why
Lab / provider	21 models; repeated releases from OpenAI, Google, Meta, Mistral, DeepSeek, and Z.ai	Descriptive only	Within-lab differences are visible, but lab, family, size, and release date still change together.
Model size	Uniform public parameter counts are unavailable	Do not compare yet	Using only open-weight parameter counts would systematically exclude or misstate closed models.
Architecture	Some public families repeat; comparable architecture metadata remains incomplete	Record first, compare later	A dense-versus-MoE result needs several documented models in each architecture group.
Release date / training cutoff	Cutoffs are not consistently disclosed	Record first, compare later	Release date is a weak proxy; training cutoff should be used only when documented consistently across the panel.

Already measured

Prompt and response sensitivity

Answer order, wording surface, evaluator context, parse rate, and stated-versus-scenario differences are the cleanest current variance candidates.

Next panel design

Repeated families matter

To test a lab or architecture effect, sample multiple dated models within each group—not one flagship per provider.

Additional axes

Same-model drift across dated releases
Prompt surface: first-person, third-person, and evaluator context
Answer-order sensitivity and response-format reliability
Model family and post-training method, where vendors document it
Inference provider, system settings, and routing changes
Language and cultural framing once multilingual items are added
Capability tier, only with a pre-specified independent capability measure

What moves together—and what we cannot yet know.

These are question themes, not traits a model “has.”

Six everyday moral themes

Two difficult trade-offs

A shared pattern, not a cause

Helping beyond one’s circle

Using harm for a larger goal

Care

Equality

Proportionality

Loyalty

Authority

Purity

Answer-order sensitivity

Wording sensitivity

Evaluator-context sensitivity

Clean response rate

Put any two measured axes together.

Which response patterns resemble one another?

Signals to inspect, not conclusions.

What this panel can and cannot test yet.

Prompt and response sensitivity

Repeated families matter